MIRA-Bench: Mirror-Induced Reasoning Ambiguity Benchmark

MIRA-Bench evaluates perception, spatial reasoning, and action decisions in reflective scenes with vision-language models. It contains 14,140 question-answer pairs over 1,011 images from 525 real-world scenes.

Hao Yin, Tianchen Guo, Heming Du, Yan Ke, Yanbin Liu, and Xin Yu.

Paper | Code

Contents

This release includes the original benchmark annotations and 430 self-collected RGB images with 430 corresponding grayscale depth images. Annotations cover 517 images from RGBD-Mirror, 64 frames from DVMD, and 430 self-collected images.

annotations/
├── questions.json
└── objects.json
images/
└── RGBD_Ben_Self_collected/
    ├── rgb/
    └── depth/

RGBD-Mirror and DVMD source images are not redistributed here. Download them from their original sources and place them under Dataset/images/RGBD-Mirror/ and Dataset/images/Published_DVMD/, preserving the split and scene structure referenced by the annotations.

• RGBD-Mirror: Depth-Aware Mirror Segmentation, CVPR 2021
• DVMD: RGB-D Video Mirror Detection, WACV 2025

Download and preparation

From the MIRA-Bench code repository, download this dataset into Dataset/:

hf download evayinhao/MIRA-Bench --repo-type dataset --local-dir Dataset

Follow the data preparation instructions to obtain the public source images and normalized depth inputs, then generate depth_colorful/, mirror_marked/, _cache/, and _cache_mm/ for evaluation.

License

The authors' original annotations and self-collected RGB/depth images are licensed under CC BY-NC 4.0 for non-commercial use; see LICENSE. RGBD-Mirror and DVMD remain subject to their original terms of use. This release does not grant rights to those third-party assets.

Citation

If you find MIRA-Bench useful in your research, please cite our paper:

@misc{yin2026depthliesbenchmarkingvisionlanguage,
  title={When Depth Lies: Benchmarking Vision-Language Models on Mirror-Induced RGB-D Ambiguity},
  author={Hao Yin and Tianchen Guo and Heming Du and Yan Ke and Yanbin Liu and Xin Yu},
  year={2026},
  eprint={2610.41651},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.41651}
}

Dataset Viewer

The default qa configuration displays all 14,140 annotation records in the test split, one record per row. The viewer uses annotations/questions.viewer.jsonl, a derived display export; the original annotations/questions.json remains the benchmark input used by the evaluation code.

The derived export preserves the original record order and values. Nested or mixed-type fields are stored as JSON text in answer_json, choices_json, slot_values_json, spatial_axes_json, target_objects_json, and related_objects_json; use json.loads to recover a present value. A missing source field appears as a null cell. source_row_index is the zero-based row index in this release, not a stable identifier across reordered releases. The source question_id values are retained as supplied and are not guaranteed to be unique.

Image and cache-path columns in this view are annotation text, not automatically decoded image columns. Some paths refer to public-source images that are not redistributed here, or to inputs generated during preparation. The value original in image_path is a preparation/evaluation marker rather than a file URL. Follow the preparation instructions above to resolve the required assets. The Viewer configuration selects only the derived QA file; objects.json remains available separately as the original object annotation file.